Over the weekend, several users on X, including reverse engineer Alexey Shabanov, reported gaining access to Grok 3, the successor to xAI's current model, Grok 2, through X's Grok chatbot app. Before their access was revoked, the users claimed to have tested the model with various queries, such as logical reasoning and coding questions.
The users noted that Grok 3 successfully answered riddles and handled requests like "Generate HTML and JavaScript code for a roulette wheel casino." However, it wasn't flawless—there were some coding mistakes, and it missed a few details in the roulette wheel code.
Shabanov mentioned that he managed to get Grok 3 to reveal its system prompt, which outlines the model's behavior instructions. One intriguing detail was that the prompt explicitly listed Donald Trump as the 47th president of the U.S., potentially addressing Grok’s previous political inaccuracies.
Grok 3, which has been in development for several months, was originally expected to launch last year but missed that deadline. Earlier this month, Musk revealed that the model had completed pre-training, a crucial development phase, and that Grok 3 was expected to release in January or early February.
xAI has been using its large data center in Memphis, which houses about 100,000 GPUs, to train Grok 3. Musk claimed in a post on X that the model was trained with “10x” more computing power than Grok 2.
Grok, designed to rival models like OpenAI's GPT-4 and Google's Gemini, can analyze images and respond to questions. It also powers several features on X, recently gaining standalone apps and potentially a "voice mode" to read out responses.
When Musk first introduced Grok two years ago, he described it as edgy, unfiltered, and anti-“woke,” willing to address controversial topics that other AI systems might avoid. While it does deliver on some of this, including responding with profanity when prompted, Grok still avoids certain political extremes. One study found that Grok tends to lean left on issues like transgender rights, diversity programs, and inequality.
Musk attributed this bias to the model’s training data, mainly sourced from public web pages, and pledged to shift Grok toward political neutrality. An "Unhinged Mode" has recently emerged, promising more controversial and offensive responses. Additionally, Musk stated that Grok 3's training data now includes court filings, enhancing its ability to understand legal topics.